After a company closes, a story forms quickly. It usually names one cause: the wrong market, raising too early, a bad hire, a co-founder split, a timing miss. The story feels earned. You were there. You watched the specific thing that did not work. The problem is not that the story is false. The problem is that it is too clean, and you are about to carry it into the next two years as though it were a precise diagnosis.
This is the overlearning trap. Not the absence of learning. The presence of too much certainty about a lesson extracted from a single complex event.
Complex failures do not come with a single cause attached
Amy Edmondson's research on organizational failure distinguishes three categories. Preventable failures involve known causes and foreseeable outcomes: a deviation from a procedure that produces an error. Complex failures involve multiple contributing factors, each necessary but none sufficient on its own. Intelligent failures produce new knowledge in conditions where the outcome could not have been predicted.
Most startup shutdowns belong in the second or third category. The market shifted, the capital environment changed, the product was right but the customer segment was wrong, two independent things went bad at the same time. The contributing factors interacted in ways that no single vantage point could see clearly, including yours.
Assigning one cause to a complex failure is itself an analytical error. The lesson that feels most true, "I raised too early," "I underestimated the sales cycle," "I picked the wrong co-founder," may identify a real factor. It rarely identifies the determining one. And building the next venture around the correction of one real-but-not-determining factor is a reliable way to repeat the underlying error from a different angle.
The mind assembles the story, it does not recall it
Here is the mechanism behind the certainty. After you know the outcome, your mind builds a coherent account of how it happened. That account draws on real memories, real data, real events. But the order, the weighting, the sense that it all pointed to the same conclusion: that is assembly, not recall.
Kahneman's work on how automatic thinking operates identifies this pattern across a large range of domains. The fast, automatic system produces narratives that cohere and feel clear. The effortful, deliberate system is what would slow the process down, test the weighting, and consider alternatives. But the deliberate system is resource-intensive and easily overridden when you already feel like you know.
After a shutdown you are depleted, which is precisely when the automatic system runs hardest and the deliberate one runs lightest. The lesson that clicks into place with the most certainty is most likely the one the automatic system found first, not the one that survived actual scrutiny. That speed is not a sign of accuracy. It is a sign of how thinly the deliberate system is running.
The pervasiveness problem
Seligman's research on how people explain negative outcomes identifies three dimensions that determine whether the explanation is adaptive or corrosive. One of them is pervasiveness: how broadly you apply the failure to your general capabilities.
"I misjudged a specific market at a specific time" is a bounded explanation. "I am bad at reading markets" is a pervasive one. "The sales cycle killed us in this vertical" is bounded. "I don't understand B2B sales" is pervasive. The difference is not modest. Pervasive explanations write off a capability entirely. Bounded ones locate the failure accurately enough to update without overwriting.
The overlearning trap tends to produce pervasive lessons because a clean lesson requires a general statement. If the real cause was a complex interaction of several factors, you cannot carry it as a single sentence. So the mind compresses it, and compression almost always increases pervasiveness. The sentence gets shorter and broader, and the capability it writes off gets larger.
What a better lesson looks like
Edmondson's framework offers a useful test: could this lesson have been known in advance, before the failure? If the answer is no, the failure is complex or intelligent, and the lesson should be probabilistic, not deterministic.
A probabilistic lesson sounds like: "In markets with long sales cycles and underfunded buyers, I now know to validate budget commitment before building." A deterministic one sounds like: "Never do enterprise sales." The first is a calibration. The second is a verdict that forecloses an entire domain based on one data point under one set of conditions.
The practical difference matters when the next decision arrives. A founder carrying a verdict stops evaluating. A founder carrying a calibration applies it conditionally, which is the right behavior for a world where the same conditions rarely repeat exactly.
A bounded, probabilistic lesson is also harder to carry. It does not close the question the way a clean verdict does. That discomfort is the signal it is more accurate. The work of the first 90 days is to resist the compression, let the failure stay complex, and extract conditions rather than verdicts. That is slower. It is also the only version that transfers correctly to what comes next.